Self-improvement proposals mined from recurring failures. Each suggests an edit to an agent's instructions or the shared team prompt. Apply runs the eval suite as a guardrail and reverts any edit that lowers the score — nothing rewrites itself without your say.